The authors augmented a multi-agent video world model with fixed-pose auxiliary cameras covering areas beyond each agent’s changing field of view.
Fixed auxiliary cameras make multi-agent world models more consistent
OverSeer jointly generates agent perspectives and stationary camera views, using the latter as a compact proxy for shared world state.
Big Tech
Runjia Li · Kam Woh Ng · Minghao Chen · Junlin Han · Jianyuan Wang · Zhaochong An · +3 more
Meta AI · University of Oxford
Research Digest··3 min read
Li and colleagues address a central problem in multi-agent video simulation: keeping separate agents’ observations consistent with one another and with earlier events.
§